AEXHQ has quietly dropped Brain v2, a minimal and extensible agent server designed for developers who want to run AI agents on their own infrastructure. Positioned explicitly as an open-source alternative to Claude Managed Agents and the OpenAI Agents SDK, Brain allows teams to deploy a single service that handles the agent loop while giving them full control over model selection and tool integration. The project hit GitHub on October 21, 2026, signaling a shift toward self-hosted, session-oriented agent architectures.

Session-Oriented Architecture

Brain’s core value proposition lies in its session-oriented primitives. Unlike ephemeral API calls, Brain creates sessions that outlive connections and server restarts. The system records an ordered history of events, allowing developers to stream agent actions live or replay them later for debugging. This architecture supports 'harnesses'β€”pluggable agent loops that manage contextβ€”and 'checkpoints' that enable portable session state. This means you can recover a session after a failure or even move it between servers during planned handoffs, a feature critical for production-grade reliability.

Tooling and Extensibility

The framework decouples the agent logic from the tools themselves. Developers can implement capabilities in their application code, attach client functions, or connect services written in any language. The documentation highlights that the model only sees the tools explicitly enabled for each session, providing a tight security boundary. For example, the included 'echo' tool runs in a separate service, demonstrating how distributed tools can be composed without bloating the core server. This flexibility extends to authentication, where each customer’s sessions and tools are isolated, ensuring multi-tenant safety.

Developer Experience and Setup

Getting started requires Go 1.26+, Node.js 24+, and an API key from a provider like OpenAI. The setup involves cloning the repository, building the 'Pi' harness, and configuring environment variables for the Brain API key and provider keys. SDKs are available for both TypeScript (@aexhq/brain) and Python (aex-brain), allowing client applications to interact with the server without needing a local Go checkout. The project also provides llms-full.txt and llms.txt files, catering specifically to coding agents that need plain-text documentation access.

Key Takeaways

  • Brain v2 is an MIT-licensed, self-hosted alternative to managed agent services like Claude Managed Agents.
  • Sessions are persistent and portable, supporting recovery and migration via checkpoints.
  • Tools are decoupled from the core server, allowing implementation in any language or service.
  • Requires Go 1.26+ and Node.js 24+ for self-hosting, with SDKs for TypeScript and Python.

The Bottom Line

Brain v2 is a smart play for developers tired of black-box agent platforms. By focusing on session persistence and model agnosticism, it offers the control we've been missing in the rush to integrate AI into products.